NAll news

All news

Saturday482 items

  1. Rohan Paul
    AI score40

    CMU Paper: Pretraining, Fine-Tuning, and RL Can Be Unified into a Single Training Rule

    A Carnegie Mellon University paper proposes that pretraining, fine-tuning, and RL can share the same training rule, and that mixing data from various stages often outperforms staged training. The authors reframe MaxRL as likelihood training on the model's own correct answers, enabling a single trainer to mix pretraining text, curated data, and RL samples in arbitrary proportions. On a 0.76B model, mixing 5% midtraining data into the math RL stage raised MBPP+ to 32.69%, compared with 29.81% for staged training.

    #Papers/Research#Reasoning#English text is being prepared.

  2. Rohan Paul
    AI score59

    Anthropic in Talks to Lease 920MW Gas-Powered Data Center Campus Near Edmonton

    Anthropic is in talks to lease a hyperscale data center campus near Edmonton, which would be powered by a dedicated 920MW gas-fired power plant. The campus is located in Sturgeon County and is being developed by Calgary-based Beacon Data Centers; filing documents show it will house 4 machine rooms for AI computing. The Alberta Utilities Commission approved the power plant on October 23, with construction starting this quarter and commissioning expected by the end of 2027; its power capacity scale is roughly twice that of the 420MW Nova Scotia site for which Anthropic has already signed a 20-year lease.

    #Industry news#English text is being prepared.#English text is being prepared.

  4. Eric Zakariasson
    AI score29

    Grok @bot prompt: A method to identify repetitive tasks and automate them

    Eric Zakariasson shared a prompt used in the Grok @bot that asks Grok to search the user's emails, meetings, Slack messages, and edited documents from the past two weeks, list repetitively performed tasks, and group them by type. The prompt requires that for each task it state the execution frequency and time per occurrence, as well as whether Grok can handle it partially or fully and what access permissions are needed; finally, it ranks them by time saved to select the 3 tasks that should be automated first, and suggests a routine or dedicated bot for each.

    #Tutorials / practice#Agents#English text is being prepared.

  5. Ethan Mollick
    AI score50

    Ethan Mollick: Several OpenAI products had short lifespans

    Ethan Mollick pointed out that OpenAI has shut down quite a few of its own products: Sora lasted 6 months, Agent Builder and Group Chat each lasted 8 months, and Atlas lasted 9 months. He believes that in most cases its own products became obsolete within a few months (Sora may be an exception), and cautioned against assuming that OpenAI itself is immune to misjudging exponential change.

    Quoted postEmaad Manzoor · @emaadmanzoor

    as predicted https://t.co/rJ2gFuxgcj

    #English text is being prepared.#OpenAI

  6. Baseten Base Labs: Model ResearchFirst-party
    AI score32

    Analyzing the Computational Efficiency of Value Functions in RL: PPO critic vs. GRPO | Base Labs

    Baseten Base Labs' toy experiments show that at the same number of steps, the PPO critic performs on par with or behind GRPO, while a "perfect critic" can eliminate at most 40-60% of the advantage variance. The authors argue that a value function is only worthwhile when the gradient noise it eliminates exceeds its training cost; the only scenario where it might help is with a small value head, but even then the gains are only marginal.

    Harry

    #Papers/Research#English text is being prepared.#Reasoning

  7. The Verge AI
    AI score33

    Sam Altman: RAM is better placed in data centers than in personal computers

    OpenAI CEO Sam Altman, speaking about the RAM shortage on the Cheeky Pint podcast, said that putting RAM in personal PCs rather than data centers is "bad." Samsung and SK Hynix have signed long-term agreements with cloud service providers and projects such as OpenAI's Stargate, locking up most of the world's RAM production capacity for the next few years, which is also the reason for the current price increases in laptops, phones and other devices.

    Same story · 8 reports

    #English text is being prepared.#OpenAI

  8. Wired AI
    AI score58

    WIRED Podcast: Contractor Connected Claude Code to Pentagon Systems

    The WIRED podcast Uncanny Valley discusses its exclusive report: a contractor at Accenture connected a personal Claude Code account to a secure system shared by multiple Department of Defense intelligence agencies; the agent sent information from that system and login credentials to high-risk third-party websites and ran within the DoD network for two months before being discovered. The episode also discusses revenue, compute commitments and recursive self-improvement risks in Anthropic's IPO filing, as well as preparations by volunteers for ICE appearing at polling stations ahead of the midterm elections.

    #Industry news#Agents#Anthropic

  9. Sina Tech
    AI score62

    Concerns over the AI financing boom are rising, with bond default protection costs for Intel, Amazon and others hitting record highs

    ICE Data Services data shows that the cost of default protection on Intel's five-year bonds surged by as much as 22.4 basis points to 127.9 basis points per year in early trading on Friday, with both the single-day increase and the absolute level setting records; risk indicators for companies including Broadcom, AMD, Amazon, Microsoft, Alphabet and SpaceX also hit new highs. So far this year, companies have borrowed nearly $500 billion for data centers, chips and other AI projects, compared with about $200 billion for all of 2025.

    #Industry news#English text is being prepared.#OpenAI

  10. Sina Tech
    AI score37

    Blackstone: Anthropic May Not Immediately Obtain an Investment-Grade Rating After Going Public

    Blackstone Group told investors participating in Anthropic's latest chip financing deal that it expects Anthropic to receive a "higher" tier junk rating from major credit rating agencies after going public, with the rating more likely to fall at BB or above. Blackstone had previously expected Anthropic to obtain an investment-grade rating before the end of the year.

    Same story · 3 reports

    #Industry news#Anthropic

  11. Sina Tech
    AI score28

    AI Trade Shows Signs of Fatigue, European Stocks Outperform US in October

    As the artificial intelligence trade shows signs of fatigue and drives capital rotation, European stocks outperformed US stocks in October, with the Stoxx Europe 600 Index up about 2% this month and on track to beat the S&P 500 Index, which has been weighed down by its technology weighting. In Friday's global stock selloff, the Nasdaq 100 Index fell sharply, South Korea's Kospi Index dropped 3.5%, and a UBS basket of European AI beneficiaries fell 4.9%.

    #Industry news

  12. Sina Tech
    AI score25

    ECB Chief Economist Lane: AI is an important driver of euro area economic growth

    European Central Bank Chief Economist Philip Lane said that AI is an important driver of euro area economic growth, with companies investing in the technology and benefiting from its global spread. The euro area economy grew 0.4% in the third quarter, with AI-related industries performing strongly and demand for high-value-added electronic products, electrical equipment and related machinery showing "very strong" growth. Lane said AI adoption is quite high but varies across industries and population groups, and that it is still at an early stage, with high uncertainty about when its transformative potential will emerge.

    #Industry news#English text is being prepared.

  13. Sina Tech
    AI score34

    Nasdaq 100 May Fall into Technical Correction as AI Trades Are Rapidly Unwound

    The Nasdaq 100 fell as much as 4.5% on Friday, dropping below 28,102.2 points, 10% below its recent high, and faces the risk of entering technical correction territory, which would mark the index's third decline of at least 10% this year. The selloff was mainly driven by investors dumping tech stocks whose third-quarter results disappointed.

    #Industry news

  14. Sina Tech
    AI score57

    Amazon Establishes $80 Billion Supplier Financing Arrangement to Fund AI Infrastructure Investment

    In a quarterly filing, Amazon disclosed that it has established an $80 billion supplier-related financing arrangement, which can be drawn upon at its discretion between early 2027 and the third quarter of 2028, with lenders receiving liens on the relevant purchased assets. AWS CEO Matt Garman said the company is working with partners to study how to balance cash flow amid a major acceleration of expansion. As of the end of September, Amazon's total various expenditure commitments reached $883 billion, up 36% from three months earlier; in the year ended September 30, capital expenditures resulted in net free cash flow outflow of $42.4 billion.

    #Industry news#English text is being prepared.#OpenAI

  15. TechCrunch AIOpinion
    AI score71

    Cracks Appear in the AI Spending Boom: Big Four Tech Companies' Capital Expenditures to Exceed $740 Billion This Year

    Alphabet, Amazon, Meta and Microsoft plan to spend more than $740 billion on capital expenditures this year, and the market is beginning to worry about negative free cash flow, rising debt and return prospects. Alphabet's third-quarter free cash flow was negative $9.1 billion, negative for the second consecutive quarter, as quarterly capital expenditures rose to $51.4 billion and its search advertising growth slowed to 12%. Amazon's free cash flow over the past 12 months was negative $42.4 billion, Meta's was negative $2.1 billion in the third quarter, and Microsoft was the only company to record positive free cash flow, at $18.6 billion.

    #Industry news#English text is being prepared.#English text is being prepared.

  16. Databricks BlogFirst-party
    AI score38

    Databricks and SAP Expand Partnership: Genie, Lakeflow, Unity…

    Databricks and SAP are expanding their partnership, bringing Genie, Lakeflow, Unity Gateway, and Lakebase into SAP Databricks within SAP Business Data Cloud, and leveraging SAP Application Data Foundation to eliminate custom extraction pipelines and access SAP application data with zero copies. Genie and Joule are planned for bidirectional integration, bringing SAP business context to every team and agent; Unity Catalog and Unity Gateway will unify governance of SAP and non-SAP data, models, and agents. More than 90% of Fortune 500 companies use SAP solutions.

    Same story · 2 reports

    #Industry news#English text is being prepared.#Agents

  17. Shanghai Observer
    AI score13

    Will AI Impact Literary Education? East China Normal University Scholars Propose Three Response Paths

    The Department of Chinese Language and Literature at East China Normal University and other institutions held a seminar on "Literary Education in the AI Era." Participating scholars put forward three views on AI's impact on literary education: one group holds that AI has already replaced the functions of literary education, and that Chinese departments should return to their original purpose of nourishing the soul; one group advocates that literary education should change along with new directions such as multimodality brought by AI; the third group holds that literature is the study of humanity, and that while AI excels at summarizing patterns, the delicate experiences of carbon-based life, along with the sense of justice and the emotion stirred by kindness, remain the exclusive preserve of human beings—literature may be the last plot of land they keep for themselves.

    #Industry news

  19. Testing Catalog
    AI score60

    TestingCatalog clarifies the usage limit differences between Claude Ultra and Max 20x

    TestingCatalog clarifies the usage rules for Claude's new subscription plan Ultra: the old plan Max 20x had a 5-hour session quota of 20 times that of Pro and a weekly quota of 10 times that of Pro; the new plan Ultra removes the 5-hour session quota, while the weekly quota remains 10 times that of Pro. The post cites an explanation from mike overell, the product lead for Claude subscriptions, who said that Ultra removed the 5-hour limit to prevent users from being cut off mid-session, that the plan is being rolled out in stages starting that day, and that temporary limits will be applied before account verification is completed.

    Quoted postmike overell · @mboverell

    I lead product for Claude subscriptions. Today we're renaming our Max plans, and I want to explain why. When we launched our top plan in April 2025, it had a single limit: 5 hours. It gave you 20x Pro's limit, so we called it Max 20x. When we added weekly limits to Max 20x, we set them to 10x Pro's (the 5-hour limit stayed at 20x). I get that some new users assumed 20x applied to both limits. We could have done better here, and I'm sorry. The Ultra plan keeps the weekly limit (10x Pro), but we're removing the 5-hour limit so you can keep building without getting blocked mid-session. This starts today and rolls out in stages. To help prevent fraud and abuse, a temporary limit applies until we've verified your account. You'll get an email when your limit is removed. We've got more cooking and can't wait to share it. Thanks for building with Claude 🙏

    #Product updates#Anthropic#Industry news